//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-23083092
// Cuda compilation tools, release 9.1, V9.1.85
// Based on LLVM 3.4svn
//

.version 6.1
.target sm_30
.address_size 64

	// .globl	_Z6oxMainv
.global .align 8 .b8 pixelID[8];
.global .align 8 .b8 resolution[8];
.global .align 4 .b8 normal[12];
.global .align 4 .b8 camPos[12];
.global .align 4 .b8 root[4];
.global .align 4 .u32 imageEnabled;
.global .texref lightmap;
.global .align 16 .b8 tileInfo[16];
.global .align 4 .u32 additive;
.global .align 1 .b8 image[1];
.global .align 1 .b8 image_HDR[1];
.global .align 1 .b8 image_HDR2[1];
.global .align 1 .b8 image_Mask[1];
.global .align 1 .b8 image_RNM0[1];
.global .align 1 .b8 image_RNM1[1];
.global .align 1 .b8 image_RNM2[1];
.global .align 1 .b8 image_RNM3[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 4 .u32 ignoreNormal;
.global .align 1 .b8 localLights[1];
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12ignoreNormalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[4] = {105, 110, 116, 0};
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 4919;
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[1];

.visible .entry _Z6oxMainv(

)
{
	.local .align 4 .b8 	__local_depot0[4];
	.reg .b64 	%SP;
	.reg .b64 	%SPL;
	.reg .pred 	%p<138>;
	.reg .b16 	%rs<162>;
	.reg .f32 	%f<1418>;
	.reg .b32 	%r<250>;
	.reg .b64 	%rd<272>;


	mov.u64 	%rd271, __local_depot0;
	cvta.local.u64 	%SP, %rd271;
	ld.global.v2.u32 	{%r30, %r31}, [pixelID];
	cvt.u64.u32	%rd10, %r30;
	cvt.u64.u32	%rd11, %r31;
	mov.u64 	%rd14, uvnormal;
	cvta.global.u64 	%rd9, %rd14;
	mov.u32 	%r28, 2;
	mov.u32 	%r29, 4;
	mov.u64 	%rd13, 0;
	// inline asm
	call (%rd8), _rt_buffer_get_64, (%rd9, %r28, %r29, %rd10, %rd11, %rd13, %rd13);
	// inline asm
	ld.u32 	%r1, [%rd8];
	shr.u32 	%r34, %r1, 16;
	cvt.u16.u32	%rs1, %r34;
	and.b16  	%rs9, %rs1, 255;
	cvt.u16.u32	%rs10, %r1;
	or.b16  	%rs11, %rs10, %rs9;
	setp.eq.s16	%p8, %rs11, 0;
	mov.f32 	%f1322, 0f00000000;
	mov.f32 	%f1323, %f1322;
	mov.f32 	%f1324, %f1322;
	@%p8 bra 	BB0_2;

	ld.u8 	%rs12, [%rd8+1];
	and.b16  	%rs14, %rs10, 255;
	cvt.rn.f32.u16	%f275, %rs14;
	div.rn.f32 	%f276, %f275, 0f437F0000;
	fma.rn.f32 	%f277, %f276, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f278, %rs12;
	div.rn.f32 	%f279, %f278, 0f437F0000;
	fma.rn.f32 	%f280, %f279, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f281, %rs9;
	div.rn.f32 	%f282, %f281, 0f437F0000;
	fma.rn.f32 	%f283, %f282, 0f40000000, 0fBF800000;
	mul.f32 	%f284, %f280, %f280;
	fma.rn.f32 	%f285, %f277, %f277, %f284;
	fma.rn.f32 	%f286, %f283, %f283, %f285;
	sqrt.rn.f32 	%f287, %f286;
	rcp.rn.f32 	%f288, %f287;
	mul.f32 	%f1322, %f277, %f288;
	mul.f32 	%f1323, %f280, %f288;
	mul.f32 	%f1324, %f283, %f288;

BB0_2:
	ld.global.v2.u32 	{%r35, %r36}, [pixelID];
	ld.global.v2.u32 	{%r38, %r39}, [tileInfo];
	add.s32 	%r2, %r35, %r38;
	add.s32 	%r3, %r36, %r39;
	setp.eq.f32	%p9, %f1323, 0f00000000;
	setp.eq.f32	%p10, %f1322, 0f00000000;
	and.pred  	%p11, %p10, %p9;
	setp.eq.f32	%p12, %f1324, 0f00000000;
	and.pred  	%p13, %p11, %p12;
	@%p13 bra 	BB0_111;
	bra.uni 	BB0_3;

BB0_111:
	ld.global.u32 	%r249, [imageEnabled];
	and.b32  	%r198, %r249, 1;
	setp.eq.b32	%p129, %r198, 1;
	@!%p129 bra 	BB0_113;
	bra.uni 	BB0_112;

BB0_112:
	cvt.u64.u32	%rd159, %r2;
	cvt.u64.u32	%rd160, %r3;
	mov.u64 	%rd163, image;
	cvta.global.u64 	%rd158, %rd163;
	// inline asm
	call (%rd157), _rt_buffer_get_64, (%rd158, %r28, %r29, %rd159, %rd160, %rd13, %rd13);
	// inline asm
	mov.u16 	%rs93, 0;
	st.v4.u8 	[%rd157], {%rs93, %rs93, %rs93, %rs93};
	ld.global.u32 	%r249, [imageEnabled];

BB0_113:
	and.b32  	%r201, %r249, 8;
	setp.eq.s32	%p130, %r201, 0;
	@%p130 bra 	BB0_115;

	cvt.u64.u32	%rd167, %r3;
	cvt.u64.u32	%rd166, %r2;
	mov.u64 	%rd170, image_Mask;
	cvta.global.u64 	%rd165, %rd170;
	// inline asm
	call (%rd164), _rt_buffer_get_64, (%rd165, %r28, %r28, %rd166, %rd167, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1283, 0f00000000;
	cvt.rzi.u32.f32	%r204, %f1283;
	cvt.u16.u32	%rs94, %r204;
	mov.u16 	%rs95, 0;
	st.v2.u8 	[%rd164], {%rs94, %rs95};
	ld.global.u32 	%r249, [imageEnabled];

BB0_115:
	cvt.u64.u32	%rd6, %r2;
	cvt.u64.u32	%rd7, %r3;
	and.b32  	%r205, %r249, 4;
	setp.eq.s32	%p131, %r205, 0;
	@%p131 bra 	BB0_119;

	ld.global.u32 	%r206, [additive];
	setp.eq.s32	%p132, %r206, 0;
	@%p132 bra 	BB0_118;

	mov.u64 	%rd183, image_HDR;
	cvta.global.u64 	%rd172, %rd183;
	mov.u32 	%r210, 8;
	// inline asm
	call (%rd171), _rt_buffer_get_64, (%rd172, %r28, %r210, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs102, %rs103, %rs104, %rs105}, [%rd171];
	// inline asm
	{  cvt.f32.f16 %f1284, %rs102;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1285, %rs103;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1286, %rs104;}

	// inline asm
	// inline asm
	call (%rd177), _rt_buffer_get_64, (%rd172, %r28, %r210, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1287, %f1284, 0f00000000;
	add.f32 	%f1288, %f1285, 0f00000000;
	add.f32 	%f1289, %f1286, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs101, %f1289;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs100, %f1288;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs99, %f1287;}

	// inline asm
	mov.u16 	%rs106, 0;
	st.v4.u16 	[%rd177], {%rs99, %rs100, %rs101, %rs106};
	bra.uni 	BB0_119;

BB0_3:
	ld.global.v2.u32 	{%r47, %r48}, [pixelID];
	cvt.u64.u32	%rd17, %r47;
	cvt.u64.u32	%rd18, %r48;
	mov.u64 	%rd26, uvpos;
	cvta.global.u64 	%rd16, %rd26;
	mov.u32 	%r44, 12;
	// inline asm
	call (%rd15), _rt_buffer_get_64, (%rd16, %r28, %r44, %rd17, %rd18, %rd13, %rd13);
	// inline asm
	ld.f32 	%f9, [%rd15+8];
	ld.f32 	%f8, [%rd15+4];
	ld.f32 	%f7, [%rd15];
	mul.f32 	%f305, %f7, 0f3456BF95;
	mul.f32 	%f306, %f8, 0f3456BF95;
	mul.f32 	%f307, %f9, 0f3456BF95;
	abs.f32 	%f308, %f1322;
	div.rn.f32 	%f309, %f305, %f308;
	abs.f32 	%f310, %f1323;
	div.rn.f32 	%f311, %f306, %f310;
	abs.f32 	%f312, %f1324;
	div.rn.f32 	%f313, %f307, %f312;
	abs.f32 	%f314, %f309;
	abs.f32 	%f315, %f311;
	abs.f32 	%f316, %f313;
	mov.f32 	%f317, 0f38D1B717;
	max.f32 	%f318, %f314, %f317;
	max.f32 	%f319, %f315, %f317;
	max.f32 	%f320, %f316, %f317;
	fma.rn.f32 	%f10, %f1322, %f318, %f7;
	fma.rn.f32 	%f11, %f1323, %f319, %f8;
	fma.rn.f32 	%f12, %f1324, %f320, %f9;
	mov.u64 	%rd27, localLights;
	cvta.global.u64 	%rd25, %rd27;
	mov.u32 	%r45, 1;
	mov.u32 	%r46, 96;
	// inline asm
	call (%rd21, %rd22, %rd23, %rd24), _rt_buffer_get_size_64, (%rd25, %r45, %r46);
	// inline asm
	cvt.u32.u64	%r4, %rd21;
	setp.eq.s32	%p14, %r4, 0;
	mov.f32 	%f1325, 0f00000000;
	mov.f32 	%f1326, %f1325;
	mov.f32 	%f1327, %f1325;
	mov.f32 	%f1328, %f1325;
	mov.f32 	%f1329, %f1325;
	mov.f32 	%f1330, %f1325;
	mov.f32 	%f1331, %f1325;
	mov.f32 	%f1332, %f1325;
	mov.f32 	%f1333, %f1325;
	mov.f32 	%f1334, %f1325;
	mov.f32 	%f1335, %f1325;
	mov.f32 	%f1336, %f1325;
	mov.f32 	%f1337, %f1325;
	mov.f32 	%f1338, %f1325;
	mov.f32 	%f1339, %f1325;
	mov.f32 	%f1340, %f1325;
	@%p14 bra 	BB0_47;

	mov.f32 	%f337, 0f40000000;
	cvt.rzi.f32.f32	%f338, %f337;
	add.f32 	%f339, %f338, %f338;
	mov.f32 	%f340, 0f40800000;
	sub.f32 	%f341, %f340, %f339;
	abs.f32 	%f13, %f341;
	mul.f32 	%f14, %f10, 0f3456BF95;
	mul.f32 	%f15, %f11, 0f3456BF95;
	mul.f32 	%f16, %f12, 0f3456BF95;
	mov.f32 	%f336, 0f00000000;
	mov.u32 	%r241, 0;
	abs.f32 	%f532, %f14;
	abs.f32 	%f533, %f15;
	max.f32 	%f534, %f532, %f533;
	abs.f32 	%f535, %f16;
	max.f32 	%f536, %f534, %f535;
	mov.f32 	%f1325, %f336;
	mov.f32 	%f1326, %f336;
	mov.f32 	%f1327, %f336;
	mov.f32 	%f1328, %f336;
	mov.f32 	%f1329, %f336;
	mov.f32 	%f1330, %f336;
	mov.f32 	%f1331, %f336;
	mov.f32 	%f1332, %f336;
	mov.f32 	%f1333, %f336;
	mov.f32 	%f1334, %f336;
	mov.f32 	%f1335, %f336;
	mov.f32 	%f1336, %f336;
	mov.f32 	%f1337, %f336;
	mov.f32 	%f1338, %f336;
	mov.f32 	%f1339, %f336;
	mov.f32 	%f1340, %f336;

BB0_5:
	cvt.u64.u32	%rd30, %r241;
	// inline asm
	call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
	// inline asm
	ld.v4.f32 	{%f344, %f345, %f346, %f347}, [%rd28+80];
	ld.v4.f32 	{%f348, %f349, %f350, %f351}, [%rd28+64];
	ld.v4.f32 	{%f352, %f353, %f354, %f355}, [%rd28+48];
	ld.v4.f32 	{%f356, %f1349, %f1350, %f359}, [%rd28+32];
	ld.v4.f32 	{%f360, %f361, %f362, %f363}, [%rd28+16];
	ld.v4.f32 	{%f364, %f365, %f366, %f367}, [%rd28];
	mov.b32 	 %r6, %f347;
	sub.f32 	%f57, %f365, %f7;
	sub.f32 	%f59, %f366, %f8;
	sub.f32 	%f61, %f367, %f9;
	mul.f32 	%f369, %f59, %f59;
	fma.rn.f32 	%f370, %f57, %f57, %f369;
	fma.rn.f32 	%f371, %f61, %f61, %f370;
	sqrt.rn.f32 	%f62, %f371;
	rcp.rn.f32 	%f372, %f62;
	mul.f32 	%f63, %f57, %f372;
	mul.f32 	%f64, %f59, %f372;
	mul.f32 	%f65, %f61, %f372;
	mul.f32 	%f66, %f62, %f363;
	abs.f32 	%f67, %f66;
	setp.lt.f32	%p15, %f67, 0f00800000;
	mul.f32 	%f373, %f67, 0f4B800000;
	selp.f32	%f374, 0fC3170000, 0fC2FE0000, %p15;
	selp.f32	%f375, %f373, %f67, %p15;
	mov.b32 	 %r54, %f375;
	and.b32  	%r55, %r54, 8388607;
	or.b32  	%r56, %r55, 1065353216;
	mov.b32 	 %f376, %r56;
	shr.u32 	%r57, %r54, 23;
	cvt.rn.f32.u32	%f377, %r57;
	add.f32 	%f378, %f374, %f377;
	setp.gt.f32	%p16, %f376, 0f3FB504F3;
	mul.f32 	%f379, %f376, 0f3F000000;
	add.f32 	%f380, %f378, 0f3F800000;
	selp.f32	%f381, %f379, %f376, %p16;
	selp.f32	%f382, %f380, %f378, %p16;
	add.f32 	%f383, %f381, 0fBF800000;
	add.f32 	%f343, %f381, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f342,%f343;
	// inline asm
	add.f32 	%f384, %f383, %f383;
	mul.f32 	%f385, %f342, %f384;
	mul.f32 	%f386, %f385, %f385;
	mov.f32 	%f387, 0f3C4CAF63;
	mov.f32 	%f388, 0f3B18F0FE;
	fma.rn.f32 	%f389, %f388, %f386, %f387;
	mov.f32 	%f390, 0f3DAAAABD;
	fma.rn.f32 	%f391, %f389, %f386, %f390;
	mul.rn.f32 	%f392, %f391, %f386;
	mul.rn.f32 	%f393, %f392, %f385;
	sub.f32 	%f394, %f383, %f385;
	neg.f32 	%f395, %f385;
	add.f32 	%f396, %f394, %f394;
	fma.rn.f32 	%f397, %f395, %f383, %f396;
	mul.rn.f32 	%f398, %f342, %f397;
	add.f32 	%f399, %f393, %f385;
	sub.f32 	%f400, %f385, %f399;
	add.f32 	%f401, %f393, %f400;
	add.f32 	%f402, %f398, %f401;
	add.f32 	%f403, %f399, %f402;
	sub.f32 	%f404, %f399, %f403;
	add.f32 	%f405, %f402, %f404;
	mov.f32 	%f406, 0f3F317200;
	mul.rn.f32 	%f407, %f382, %f406;
	mov.f32 	%f408, 0f35BFBE8E;
	mul.rn.f32 	%f409, %f382, %f408;
	add.f32 	%f410, %f407, %f403;
	sub.f32 	%f411, %f407, %f410;
	add.f32 	%f412, %f403, %f411;
	add.f32 	%f413, %f405, %f412;
	add.f32 	%f414, %f409, %f413;
	add.f32 	%f415, %f410, %f414;
	sub.f32 	%f416, %f410, %f415;
	add.f32 	%f417, %f414, %f416;
	mul.rn.f32 	%f68, %f340, %f415;
	neg.f32 	%f419, %f68;
	fma.rn.f32 	%f420, %f340, %f415, %f419;
	fma.rn.f32 	%f421, %f340, %f417, %f420;
	fma.rn.f32 	%f69, %f336, %f415, %f421;
	add.rn.f32 	%f70, %f68, %f69;
	mov.b32 	 %r58, %f70;
	setp.eq.s32	%p1, %r58, 1118925336;
	add.s32 	%r59, %r58, -1;
	mov.b32 	 %f423, %r59;
	selp.f32	%f424, %f423, %f70, %p1;
	mul.f32 	%f425, %f424, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f426, %f425;
	mov.f32 	%f427, 0fBF317200;
	fma.rn.f32 	%f428, %f426, %f427, %f424;
	mov.f32 	%f429, 0fB5BFBE8E;
	fma.rn.f32 	%f430, %f426, %f429, %f428;
	mul.f32 	%f431, %f430, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f432, %f431;
	add.f32 	%f433, %f426, 0f00000000;
	ex2.approx.f32 	%f434, %f433;
	mul.f32 	%f435, %f432, %f434;
	setp.lt.f32	%p17, %f424, 0fC2D20000;
	selp.f32	%f436, 0f00000000, %f435, %p17;
	setp.gt.f32	%p18, %f424, 0f42D20000;
	selp.f32	%f1341, 0f7F800000, %f436, %p18;
	setp.eq.f32	%p19, %f1341, 0f7F800000;
	@%p19 bra 	BB0_7;

	neg.f32 	%f437, %f70;
	add.rn.f32 	%f438, %f68, %f437;
	add.rn.f32 	%f439, %f438, %f69;
	add.f32 	%f440, %f439, 0f37000000;
	selp.f32	%f441, %f440, %f439, %p1;
	fma.rn.f32 	%f1341, %f1341, %f441, %f1341;

BB0_7:
	setp.lt.f32	%p20, %f66, 0f00000000;
	setp.eq.f32	%p21, %f13, 0f3F800000;
	and.pred  	%p2, %p20, %p21;
	mov.b32 	 %r60, %f1341;
	xor.b32  	%r61, %r60, -2147483648;
	mov.b32 	 %f442, %r61;
	selp.f32	%f1343, %f442, %f1341, %p2;
	setp.eq.f32	%p22, %f66, 0f00000000;
	@%p22 bra 	BB0_10;
	bra.uni 	BB0_8;

BB0_10:
	add.f32 	%f445, %f66, %f66;
	selp.f32	%f1343, %f445, 0f00000000, %p21;
	bra.uni 	BB0_11;

BB0_8:
	setp.geu.f32	%p23, %f66, 0f00000000;
	@%p23 bra 	BB0_11;

	cvt.rzi.f32.f32	%f444, %f340;
	setp.neu.f32	%p24, %f444, 0f40800000;
	selp.f32	%f1343, 0f7FFFFFFF, %f1343, %p24;

BB0_11:
	add.f32 	%f446, %f67, 0f40800000;
	mov.b32 	 %r62, %f446;
	setp.lt.s32	%p26, %r62, 2139095040;
	@%p26 bra 	BB0_16;

	setp.gtu.f32	%p27, %f67, 0f7F800000;
	@%p27 bra 	BB0_15;
	bra.uni 	BB0_13;

BB0_15:
	add.f32 	%f1343, %f66, 0f40800000;
	bra.uni 	BB0_16;

BB0_13:
	setp.neu.f32	%p28, %f67, 0f7F800000;
	@%p28 bra 	BB0_16;

	selp.f32	%f1343, 0fFF800000, 0f7F800000, %p2;

BB0_16:
	mul.f32 	%f447, %f62, %f361;
	mov.f32 	%f1372, 0f3F800000;
	sub.f32 	%f449, %f1372, %f1343;
	setp.eq.f32	%p29, %f66, 0f3F800000;
	selp.f32	%f450, 0f00000000, %f449, %p29;
	cvt.sat.f32.f32	%f451, %f450;
	fma.rn.f32 	%f452, %f447, %f447, %f362;
	div.rn.f32 	%f96, %f451, %f452;
	mul.f32 	%f453, %f1323, %f64;
	fma.rn.f32 	%f454, %f1322, %f63, %f453;
	fma.rn.f32 	%f97, %f1324, %f65, %f454;
	setp.eq.f32	%p30, %f364, 0f3F800000;
	@%p30 bra 	BB0_23;
	bra.uni 	BB0_17;

BB0_23:
	setp.leu.f32	%p34, %f359, 0f00000000;
	@%p34 bra 	BB0_19;

	setp.geu.f32	%p35, %f355, 0f00000000;
	@%p35 bra 	BB0_25;

	sub.f32 	%f1321, %f367, %f9;
	sub.f32 	%f1320, %f365, %f7;
	sub.f32 	%f1319, %f366, %f8;
	mul.f32 	%f477, %f345, %f1319;
	fma.rn.f32 	%f478, %f344, %f1320, %f477;
	fma.rn.f32 	%f479, %f346, %f1321, %f478;
	rcp.rn.f32 	%f480, %f479;
	mul.f32 	%f1344, %f1320, %f480;
	mul.f32 	%f1345, %f1319, %f480;
	mul.f32 	%f1346, %f1321, %f480;
	neg.f32 	%f355, %f355;
	bra.uni 	BB0_27;

BB0_17:
	setp.eq.f32	%p31, %f364, 0f40000000;
	@%p31 bra 	BB0_21;
	bra.uni 	BB0_18;

BB0_21:
	setp.leu.f32	%p33, %f359, 0f00000000;
	@%p33 bra 	BB0_19;

	mul.f32 	%f471, %f353, %f64;
	fma.rn.f32 	%f472, %f352, %f63, %f471;
	mul.f32 	%f473, %f349, %f64;
	fma.rn.f32 	%f474, %f348, %f63, %f473;
	mul.f32 	%f475, %f345, %f64;
	fma.rn.f32 	%f476, %f344, %f63, %f475;
	fma.rn.f32 	%f468, %f354, %f65, %f472;
	fma.rn.f32 	%f469, %f350, %f65, %f474;
	fma.rn.f32 	%f470, %f346, %f65, %f476;
	cvt.rzi.s32.f32	%r63, %f359;
	mov.u32 	%r64, 6;
	mov.u32 	%r65, 0;
	// inline asm
	call (%f464, %f465, %f466, %f467), _rt_texture_get_base_id, (%r63, %r64, %f468, %f469, %f470, %r65);
	// inline asm
	mul.f32 	%f1348, %f356, %f464;
	mul.f32 	%f1349, %f1349, %f465;
	mul.f32 	%f1350, %f1350, %f466;
	bra.uni 	BB0_28;

BB0_18:
	setp.neu.f32	%p32, %f364, 0f40800000;
	@%p32 bra 	BB0_19;

	mul.f32 	%f455, %f344, %f63;
	mul.f32 	%f456, %f345, %f64;
	neg.f32 	%f457, %f456;
	sub.f32 	%f458, %f457, %f455;
	mul.f32 	%f459, %f346, %f65;
	sub.f32 	%f460, %f458, %f459;
	fma.rn.f32 	%f461, %f359, %f460, %f355;
	cvt.sat.f32.f32	%f462, %f461;
	mul.f32 	%f463, %f462, %f462;
	mul.f32 	%f1351, %f96, %f463;
	mov.f32 	%f1348, %f356;
	bra.uni 	BB0_29;

BB0_19:
	mov.f32 	%f1348, %f356;
	bra.uni 	BB0_28;

BB0_25:
	mov.f32 	%f1344, %f63;
	mov.f32 	%f1345, %f64;
	mov.f32 	%f1346, %f65;

BB0_27:
	mul.f32 	%f489, %f346, %f65;
	mul.f32 	%f490, %f345, %f64;
	neg.f32 	%f491, %f490;
	mul.f32 	%f492, %f344, %f63;
	sub.f32 	%f493, %f491, %f492;
	sub.f32 	%f494, %f493, %f489;
	setp.gt.f32	%p36, %f494, 0f00000000;
	selp.f32	%f495, 0f3F800000, 0f00000000, %p36;
	mul.f32 	%f496, %f353, %f1345;
	fma.rn.f32 	%f497, %f352, %f1344, %f496;
	mul.f32 	%f498, %f349, %f1345;
	fma.rn.f32 	%f499, %f348, %f1344, %f498;
	fma.rn.f32 	%f500, %f354, %f1346, %f497;
	fma.rn.f32 	%f501, %f350, %f1346, %f499;
	fma.rn.f32 	%f502, %f355, %f500, 0f3F000000;
	mov.f32 	%f503, 0f3F800000;
	sub.f32 	%f485, %f503, %f502;
	fma.rn.f32 	%f486, %f355, %f501, 0f3F000000;
	cvt.rzi.s32.f32	%r66, %f359;
	mov.f32 	%f488, 0f00000000;
	// inline asm
	call (%f481, %f482, %f483, %f484), _rt_texture_get_f_id, (%r66, %r28, %f485, %f486, %f488, %f488);
	// inline asm
	mul.f32 	%f504, %f495, %f481;
	mul.f32 	%f505, %f495, %f482;
	mul.f32 	%f506, %f495, %f483;
	mul.f32 	%f1348, %f356, %f504;
	mul.f32 	%f1349, %f1349, %f505;
	mul.f32 	%f1350, %f1350, %f506;

BB0_28:
	mov.f32 	%f1351, %f96;

BB0_29:
	max.f32 	%f522, %f1348, %f1349;
	max.f32 	%f523, %f522, %f1350;
	mul.f32 	%f524, %f1351, %f523;
	setp.lt.f32	%p38, %f524, 0f3727C5AC;
	mov.pred 	%p137, -1;
	mov.f32 	%f132, 0f00000000;
	mov.f32 	%f1353, %f132;
	mov.f32 	%f1354, %f132;
	mov.f32 	%f1355, %f132;
	mov.f32 	%f1356, %f132;
	mov.f32 	%f1357, %f132;
	mov.f32 	%f1358, %f132;
	mov.f32 	%f1359, %f132;
	mov.f32 	%f1360, %f132;
	mov.f32 	%f1361, %f132;
	mov.f32 	%f1362, %f132;
	mov.f32 	%f1363, %f132;
	mov.f32 	%f1364, %f132;
	mov.f32 	%f1365, %f132;
	mov.f32 	%f1366, %f132;
	@%p38 bra 	BB0_31;

	ld.global.u32 	%r68, [ignoreNormal];
	setp.eq.s32	%p40, %r68, 0;
	selp.f32	%f525, %f97, 0f3F800000, %p40;
	cvt.sat.f32.f32	%f526, %f525;
	mul.f32 	%f527, %f1351, %f526;
	mul.f32 	%f132, %f1348, %f527;
	mul.f32 	%f1353, %f1349, %f527;
	mul.f32 	%f1354, %f1350, %f527;
	mul.f32 	%f528, %f1351, 0f3E800000;
	mul.f32 	%f1355, %f1348, %f528;
	mul.f32 	%f1356, %f1349, %f528;
	mul.f32 	%f1357, %f1350, %f528;
	mul.f32 	%f1358, %f63, %f1355;
	mul.f32 	%f1359, %f63, %f1356;
	mul.f32 	%f1360, %f63, %f1357;
	mul.f32 	%f1361, %f64, %f1355;
	mul.f32 	%f1362, %f64, %f1356;
	mul.f32 	%f1363, %f64, %f1357;
	mul.f32 	%f1364, %f65, %f1355;
	mul.f32 	%f1365, %f65, %f1356;
	mul.f32 	%f1366, %f65, %f1357;
	mov.pred 	%p137, 0;

BB0_31:
	@%p137 bra 	BB0_46;

	setp.eq.s32	%p41, %r6, 0;
	mov.u16 	%rs161, 0;
	@%p41 bra 	BB0_43;

	abs.s32 	%r8, %r6;
	mov.f32 	%f1371, 0f00000000;
	setp.lt.s32	%p42, %r8, 1;
	@%p42 bra 	BB0_42;

	max.f32 	%f148, %f536, %f317;
	and.b32  	%r9, %r8, 3;
	setp.eq.s32	%p43, %r9, 0;
	add.u64 	%rd35, %SP, 0;
	cvta.to.local.u64 	%rd2, %rd35;
	mov.f32 	%f1371, 0f00000000;
	mov.u32 	%r245, 0;
	@%p43 bra 	BB0_40;

	setp.eq.s32	%p44, %r9, 1;
	mov.f32 	%f1368, 0f00000000;
	mov.u32 	%r243, 0;
	@%p44 bra 	BB0_39;

	setp.eq.s32	%p45, %r9, 2;
	mov.f32 	%f1367, 0f00000000;
	mov.u32 	%r242, 0;
	@%p45 bra 	BB0_38;

	sub.f32 	%f548, %f365, %f360;
	sub.f32 	%f549, %f366, %f360;
	sub.f32 	%f550, %f367, %f360;
	sub.f32 	%f551, %f548, %f7;
	sub.f32 	%f552, %f549, %f8;
	sub.f32 	%f553, %f550, %f9;
	mul.f32 	%f554, %f552, %f552;
	fma.rn.f32 	%f555, %f551, %f551, %f554;
	fma.rn.f32 	%f556, %f553, %f553, %f555;
	sqrt.rn.f32 	%f547, %f556;
	rcp.rn.f32 	%f557, %f547;
	mul.f32 	%f543, %f557, %f551;
	mul.f32 	%f544, %f557, %f552;
	mul.f32 	%f545, %f557, %f553;
	ld.global.u32 	%r76, [imageEnabled];
	and.b32  	%r77, %r76, 32;
	setp.eq.s32	%p46, %r77, 0;
	selp.f32	%f558, 0f3F800000, 0f41200000, %p46;
	mul.f32 	%f546, %f558, %f148;
	mov.u32 	%r78, 1065353216;
	st.local.u32 	[%rd2], %r78;
	ld.global.u32 	%r72, [root];
	// inline asm
	call _rt_trace_64, (%r72, %f10, %f11, %f12, %f543, %f544, %f545, %r45, %f546, %f547, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f559, [%rd2];
	add.f32 	%f1367, %f559, 0f00000000;
	mov.u32 	%r242, %r45;

BB0_38:
	cvt.rn.f32.s32	%f568, %r242;
	mul.f32 	%f569, %f568, 0f3DD32618;
	cvt.rmi.f32.f32	%f570, %f569;
	sub.f32 	%f571, %f569, %f570;
	mul.f32 	%f572, %f568, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f573, %f572;
	sub.f32 	%f574, %f572, %f573;
	mul.f32 	%f575, %f568, 0f3DC74539;
	cvt.rmi.f32.f32	%f576, %f575;
	sub.f32 	%f577, %f575, %f576;
	add.f32 	%f578, %f574, 0f4199851F;
	add.f32 	%f579, %f577, 0f4199851F;
	add.f32 	%f580, %f571, 0f4199851F;
	mul.f32 	%f581, %f574, %f579;
	fma.rn.f32 	%f582, %f571, %f578, %f581;
	fma.rn.f32 	%f583, %f580, %f577, %f582;
	add.f32 	%f584, %f571, %f583;
	add.f32 	%f585, %f574, %f583;
	add.f32 	%f586, %f577, %f583;
	add.f32 	%f587, %f584, %f585;
	mul.f32 	%f588, %f586, %f587;
	cvt.rmi.f32.f32	%f589, %f588;
	sub.f32 	%f590, %f588, %f589;
	add.f32 	%f591, %f584, %f586;
	mul.f32 	%f592, %f585, %f591;
	cvt.rmi.f32.f32	%f593, %f592;
	sub.f32 	%f594, %f592, %f593;
	add.f32 	%f595, %f585, %f586;
	mul.f32 	%f596, %f584, %f595;
	cvt.rmi.f32.f32	%f597, %f596;
	sub.f32 	%f598, %f596, %f597;
	fma.rn.f32 	%f599, %f590, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f600, %f594, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f601, %f598, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f602, %f360, %f599, %f365;
	fma.rn.f32 	%f603, %f360, %f600, %f366;
	fma.rn.f32 	%f604, %f360, %f601, %f367;
	sub.f32 	%f605, %f602, %f7;
	sub.f32 	%f606, %f603, %f8;
	sub.f32 	%f607, %f604, %f9;
	mul.f32 	%f608, %f606, %f606;
	fma.rn.f32 	%f609, %f605, %f605, %f608;
	fma.rn.f32 	%f610, %f607, %f607, %f609;
	sqrt.rn.f32 	%f567, %f610;
	rcp.rn.f32 	%f611, %f567;
	mul.f32 	%f563, %f611, %f605;
	mul.f32 	%f564, %f611, %f606;
	mul.f32 	%f565, %f611, %f607;
	ld.global.u32 	%r82, [imageEnabled];
	and.b32  	%r83, %r82, 32;
	setp.eq.s32	%p47, %r83, 0;
	selp.f32	%f612, 0f3F800000, 0f41200000, %p47;
	mul.f32 	%f566, %f612, %f148;
	mov.u32 	%r84, 1065353216;
	st.local.u32 	[%rd2], %r84;
	ld.global.u32 	%r79, [root];
	// inline asm
	call _rt_trace_64, (%r79, %f10, %f11, %f12, %f563, %f564, %f565, %r45, %f566, %f567, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f613, [%rd2];
	add.f32 	%f1368, %f1367, %f613;
	add.s32 	%r243, %r242, 1;

BB0_39:
	cvt.rn.f32.s32	%f622, %r243;
	mul.f32 	%f623, %f622, 0f3DD32618;
	cvt.rmi.f32.f32	%f624, %f623;
	sub.f32 	%f625, %f623, %f624;
	mul.f32 	%f626, %f622, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f627, %f626;
	sub.f32 	%f628, %f626, %f627;
	mul.f32 	%f629, %f622, 0f3DC74539;
	cvt.rmi.f32.f32	%f630, %f629;
	sub.f32 	%f631, %f629, %f630;
	add.f32 	%f632, %f628, 0f4199851F;
	add.f32 	%f633, %f631, 0f4199851F;
	add.f32 	%f634, %f625, 0f4199851F;
	mul.f32 	%f635, %f628, %f633;
	fma.rn.f32 	%f636, %f625, %f632, %f635;
	fma.rn.f32 	%f637, %f634, %f631, %f636;
	add.f32 	%f638, %f625, %f637;
	add.f32 	%f639, %f628, %f637;
	add.f32 	%f640, %f631, %f637;
	add.f32 	%f641, %f638, %f639;
	mul.f32 	%f642, %f640, %f641;
	cvt.rmi.f32.f32	%f643, %f642;
	sub.f32 	%f644, %f642, %f643;
	add.f32 	%f645, %f638, %f640;
	mul.f32 	%f646, %f639, %f645;
	cvt.rmi.f32.f32	%f647, %f646;
	sub.f32 	%f648, %f646, %f647;
	add.f32 	%f649, %f639, %f640;
	mul.f32 	%f650, %f638, %f649;
	cvt.rmi.f32.f32	%f651, %f650;
	sub.f32 	%f652, %f650, %f651;
	fma.rn.f32 	%f653, %f644, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f654, %f648, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f655, %f652, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f656, %f360, %f653, %f365;
	fma.rn.f32 	%f657, %f360, %f654, %f366;
	fma.rn.f32 	%f658, %f360, %f655, %f367;
	sub.f32 	%f659, %f656, %f7;
	sub.f32 	%f660, %f657, %f8;
	sub.f32 	%f661, %f658, %f9;
	mul.f32 	%f662, %f660, %f660;
	fma.rn.f32 	%f663, %f659, %f659, %f662;
	fma.rn.f32 	%f664, %f661, %f661, %f663;
	sqrt.rn.f32 	%f621, %f664;
	rcp.rn.f32 	%f665, %f621;
	mul.f32 	%f617, %f665, %f659;
	mul.f32 	%f618, %f665, %f660;
	mul.f32 	%f619, %f665, %f661;
	ld.global.u32 	%r88, [imageEnabled];
	and.b32  	%r89, %r88, 32;
	setp.eq.s32	%p48, %r89, 0;
	selp.f32	%f666, 0f3F800000, 0f41200000, %p48;
	mul.f32 	%f620, %f666, %f148;
	mov.u32 	%r90, 1065353216;
	st.local.u32 	[%rd2], %r90;
	ld.global.u32 	%r85, [root];
	mov.u32 	%r86, 1;
	// inline asm
	call _rt_trace_64, (%r85, %f10, %f11, %f12, %f617, %f618, %f619, %r86, %f620, %f621, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f667, [%rd2];
	add.f32 	%f1371, %f1368, %f667;
	add.s32 	%r245, %r243, 1;

BB0_40:
	setp.lt.u32	%p49, %r8, 4;
	@%p49 bra 	BB0_42;

BB0_41:
	cvt.rn.f32.s32	%f700, %r245;
	mul.f32 	%f701, %f700, 0f3DD32618;
	cvt.rmi.f32.f32	%f702, %f701;
	sub.f32 	%f703, %f701, %f702;
	mul.f32 	%f704, %f700, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f705, %f704;
	sub.f32 	%f706, %f704, %f705;
	mul.f32 	%f707, %f700, 0f3DC74539;
	cvt.rmi.f32.f32	%f708, %f707;
	sub.f32 	%f709, %f707, %f708;
	add.f32 	%f710, %f706, 0f4199851F;
	add.f32 	%f711, %f709, 0f4199851F;
	add.f32 	%f712, %f703, 0f4199851F;
	mul.f32 	%f713, %f706, %f711;
	fma.rn.f32 	%f714, %f703, %f710, %f713;
	fma.rn.f32 	%f715, %f712, %f709, %f714;
	add.f32 	%f716, %f703, %f715;
	add.f32 	%f717, %f706, %f715;
	add.f32 	%f718, %f709, %f715;
	add.f32 	%f719, %f716, %f717;
	mul.f32 	%f720, %f718, %f719;
	cvt.rmi.f32.f32	%f721, %f720;
	sub.f32 	%f722, %f720, %f721;
	add.f32 	%f723, %f716, %f718;
	mul.f32 	%f724, %f717, %f723;
	cvt.rmi.f32.f32	%f725, %f724;
	sub.f32 	%f726, %f724, %f725;
	add.f32 	%f727, %f717, %f718;
	mul.f32 	%f728, %f716, %f727;
	cvt.rmi.f32.f32	%f729, %f728;
	sub.f32 	%f730, %f728, %f729;
	fma.rn.f32 	%f731, %f722, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f732, %f726, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f733, %f730, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f734, %f360, %f731, %f365;
	fma.rn.f32 	%f735, %f360, %f732, %f366;
	fma.rn.f32 	%f736, %f360, %f733, %f367;
	sub.f32 	%f737, %f734, %f7;
	sub.f32 	%f738, %f735, %f8;
	sub.f32 	%f739, %f736, %f9;
	mul.f32 	%f740, %f738, %f738;
	fma.rn.f32 	%f741, %f737, %f737, %f740;
	fma.rn.f32 	%f742, %f739, %f739, %f741;
	sqrt.rn.f32 	%f675, %f742;
	rcp.rn.f32 	%f743, %f675;
	mul.f32 	%f671, %f743, %f737;
	mul.f32 	%f672, %f743, %f738;
	mul.f32 	%f673, %f743, %f739;
	ld.global.u32 	%r103, [imageEnabled];
	and.b32  	%r104, %r103, 32;
	setp.eq.s32	%p50, %r104, 0;
	selp.f32	%f744, 0f3F800000, 0f41200000, %p50;
	mul.f32 	%f674, %f744, %f148;
	mov.u32 	%r105, 1065353216;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r91, [root];
	mov.u32 	%r101, 1;
	// inline asm
	call _rt_trace_64, (%r91, %f10, %f11, %f12, %f671, %f672, %f673, %r101, %f674, %f675, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f745, [%rd2];
	add.f32 	%f746, %f1371, %f745;
	add.s32 	%r106, %r245, 1;
	cvt.rn.f32.s32	%f747, %r106;
	mul.f32 	%f748, %f747, 0f3DD32618;
	cvt.rmi.f32.f32	%f749, %f748;
	sub.f32 	%f750, %f748, %f749;
	mul.f32 	%f751, %f747, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f752, %f751;
	sub.f32 	%f753, %f751, %f752;
	mul.f32 	%f754, %f747, 0f3DC74539;
	cvt.rmi.f32.f32	%f755, %f754;
	sub.f32 	%f756, %f754, %f755;
	add.f32 	%f757, %f753, 0f4199851F;
	add.f32 	%f758, %f756, 0f4199851F;
	add.f32 	%f759, %f750, 0f4199851F;
	mul.f32 	%f760, %f753, %f758;
	fma.rn.f32 	%f761, %f750, %f757, %f760;
	fma.rn.f32 	%f762, %f759, %f756, %f761;
	add.f32 	%f763, %f750, %f762;
	add.f32 	%f764, %f753, %f762;
	add.f32 	%f765, %f756, %f762;
	add.f32 	%f766, %f763, %f764;
	mul.f32 	%f767, %f765, %f766;
	cvt.rmi.f32.f32	%f768, %f767;
	sub.f32 	%f769, %f767, %f768;
	add.f32 	%f770, %f763, %f765;
	mul.f32 	%f771, %f764, %f770;
	cvt.rmi.f32.f32	%f772, %f771;
	sub.f32 	%f773, %f771, %f772;
	add.f32 	%f774, %f764, %f765;
	mul.f32 	%f775, %f763, %f774;
	cvt.rmi.f32.f32	%f776, %f775;
	sub.f32 	%f777, %f775, %f776;
	fma.rn.f32 	%f778, %f769, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f779, %f773, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f780, %f777, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f781, %f360, %f778, %f365;
	fma.rn.f32 	%f782, %f360, %f779, %f366;
	fma.rn.f32 	%f783, %f360, %f780, %f367;
	sub.f32 	%f784, %f781, %f7;
	sub.f32 	%f785, %f782, %f8;
	sub.f32 	%f786, %f783, %f9;
	mul.f32 	%f787, %f785, %f785;
	fma.rn.f32 	%f788, %f784, %f784, %f787;
	fma.rn.f32 	%f789, %f786, %f786, %f788;
	sqrt.rn.f32 	%f683, %f789;
	rcp.rn.f32 	%f790, %f683;
	mul.f32 	%f679, %f790, %f784;
	mul.f32 	%f680, %f790, %f785;
	mul.f32 	%f681, %f790, %f786;
	ld.global.u32 	%r107, [imageEnabled];
	and.b32  	%r108, %r107, 32;
	setp.eq.s32	%p51, %r108, 0;
	selp.f32	%f791, 0f3F800000, 0f41200000, %p51;
	mul.f32 	%f682, %f791, %f148;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r94, [root];
	// inline asm
	call _rt_trace_64, (%r94, %f10, %f11, %f12, %f679, %f680, %f681, %r101, %f682, %f683, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f792, [%rd2];
	add.f32 	%f793, %f746, %f792;
	add.s32 	%r109, %r245, 2;
	cvt.rn.f32.s32	%f794, %r109;
	mul.f32 	%f795, %f794, 0f3DD32618;
	cvt.rmi.f32.f32	%f796, %f795;
	sub.f32 	%f797, %f795, %f796;
	mul.f32 	%f798, %f794, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f799, %f798;
	sub.f32 	%f800, %f798, %f799;
	mul.f32 	%f801, %f794, 0f3DC74539;
	cvt.rmi.f32.f32	%f802, %f801;
	sub.f32 	%f803, %f801, %f802;
	add.f32 	%f804, %f800, 0f4199851F;
	add.f32 	%f805, %f803, 0f4199851F;
	add.f32 	%f806, %f797, 0f4199851F;
	mul.f32 	%f807, %f800, %f805;
	fma.rn.f32 	%f808, %f797, %f804, %f807;
	fma.rn.f32 	%f809, %f806, %f803, %f808;
	add.f32 	%f810, %f797, %f809;
	add.f32 	%f811, %f800, %f809;
	add.f32 	%f812, %f803, %f809;
	add.f32 	%f813, %f810, %f811;
	mul.f32 	%f814, %f812, %f813;
	cvt.rmi.f32.f32	%f815, %f814;
	sub.f32 	%f816, %f814, %f815;
	add.f32 	%f817, %f810, %f812;
	mul.f32 	%f818, %f811, %f817;
	cvt.rmi.f32.f32	%f819, %f818;
	sub.f32 	%f820, %f818, %f819;
	add.f32 	%f821, %f811, %f812;
	mul.f32 	%f822, %f810, %f821;
	cvt.rmi.f32.f32	%f823, %f822;
	sub.f32 	%f824, %f822, %f823;
	fma.rn.f32 	%f825, %f816, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f826, %f820, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f827, %f824, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f828, %f360, %f825, %f365;
	fma.rn.f32 	%f829, %f360, %f826, %f366;
	fma.rn.f32 	%f830, %f360, %f827, %f367;
	sub.f32 	%f831, %f828, %f7;
	sub.f32 	%f832, %f829, %f8;
	sub.f32 	%f833, %f830, %f9;
	mul.f32 	%f834, %f832, %f832;
	fma.rn.f32 	%f835, %f831, %f831, %f834;
	fma.rn.f32 	%f836, %f833, %f833, %f835;
	sqrt.rn.f32 	%f691, %f836;
	rcp.rn.f32 	%f837, %f691;
	mul.f32 	%f687, %f837, %f831;
	mul.f32 	%f688, %f837, %f832;
	mul.f32 	%f689, %f837, %f833;
	ld.global.u32 	%r110, [imageEnabled];
	and.b32  	%r111, %r110, 32;
	setp.eq.s32	%p52, %r111, 0;
	selp.f32	%f838, 0f3F800000, 0f41200000, %p52;
	mul.f32 	%f690, %f838, %f148;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r97, [root];
	// inline asm
	call _rt_trace_64, (%r97, %f10, %f11, %f12, %f687, %f688, %f689, %r101, %f690, %f691, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f839, [%rd2];
	add.f32 	%f840, %f793, %f839;
	add.s32 	%r112, %r245, 3;
	cvt.rn.f32.s32	%f841, %r112;
	mul.f32 	%f842, %f841, 0f3DD32618;
	cvt.rmi.f32.f32	%f843, %f842;
	sub.f32 	%f844, %f842, %f843;
	mul.f32 	%f845, %f841, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f846, %f845;
	sub.f32 	%f847, %f845, %f846;
	mul.f32 	%f848, %f841, 0f3DC74539;
	cvt.rmi.f32.f32	%f849, %f848;
	sub.f32 	%f850, %f848, %f849;
	add.f32 	%f851, %f847, 0f4199851F;
	add.f32 	%f852, %f850, 0f4199851F;
	add.f32 	%f853, %f844, 0f4199851F;
	mul.f32 	%f854, %f847, %f852;
	fma.rn.f32 	%f855, %f844, %f851, %f854;
	fma.rn.f32 	%f856, %f853, %f850, %f855;
	add.f32 	%f857, %f844, %f856;
	add.f32 	%f858, %f847, %f856;
	add.f32 	%f859, %f850, %f856;
	add.f32 	%f860, %f857, %f858;
	mul.f32 	%f861, %f859, %f860;
	cvt.rmi.f32.f32	%f862, %f861;
	sub.f32 	%f863, %f861, %f862;
	add.f32 	%f864, %f857, %f859;
	mul.f32 	%f865, %f858, %f864;
	cvt.rmi.f32.f32	%f866, %f865;
	sub.f32 	%f867, %f865, %f866;
	add.f32 	%f868, %f858, %f859;
	mul.f32 	%f869, %f857, %f868;
	cvt.rmi.f32.f32	%f870, %f869;
	sub.f32 	%f871, %f869, %f870;
	fma.rn.f32 	%f872, %f863, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f873, %f867, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f874, %f871, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f875, %f360, %f872, %f365;
	fma.rn.f32 	%f876, %f360, %f873, %f366;
	fma.rn.f32 	%f877, %f360, %f874, %f367;
	sub.f32 	%f878, %f875, %f7;
	sub.f32 	%f879, %f876, %f8;
	sub.f32 	%f880, %f877, %f9;
	mul.f32 	%f881, %f879, %f879;
	fma.rn.f32 	%f882, %f878, %f878, %f881;
	fma.rn.f32 	%f883, %f880, %f880, %f882;
	sqrt.rn.f32 	%f699, %f883;
	rcp.rn.f32 	%f884, %f699;
	mul.f32 	%f695, %f884, %f878;
	mul.f32 	%f696, %f884, %f879;
	mul.f32 	%f697, %f884, %f880;
	ld.global.u32 	%r113, [imageEnabled];
	and.b32  	%r114, %r113, 32;
	setp.eq.s32	%p53, %r114, 0;
	selp.f32	%f885, 0f3F800000, 0f41200000, %p53;
	mul.f32 	%f698, %f885, %f148;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r100, [root];
	// inline asm
	call _rt_trace_64, (%r100, %f10, %f11, %f12, %f695, %f696, %f697, %r101, %f698, %f699, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f886, [%rd2];
	add.f32 	%f1371, %f840, %f886;
	add.s32 	%r245, %r245, 4;
	setp.lt.s32	%p54, %r245, %r8;
	@%p54 bra 	BB0_41;

BB0_42:
	cvt.rn.f32.s32	%f887, %r8;
	div.rn.f32 	%f1372, %f1371, %f887;
	shr.u32 	%r115, %r6, 31;
	cvt.u16.u32	%rs161, %r115;

BB0_43:
	fma.rn.f32 	%f1340, %f132, %f1372, %f1340;
	fma.rn.f32 	%f1339, %f1353, %f1372, %f1339;
	fma.rn.f32 	%f1338, %f1354, %f1372, %f1338;
	fma.rn.f32 	%f1337, %f1355, %f1372, %f1337;
	fma.rn.f32 	%f1336, %f1356, %f1372, %f1336;
	fma.rn.f32 	%f1335, %f1357, %f1372, %f1335;
	fma.rn.f32 	%f1334, %f1358, %f1372, %f1334;
	fma.rn.f32 	%f1333, %f1359, %f1372, %f1333;
	fma.rn.f32 	%f1332, %f1360, %f1372, %f1332;
	fma.rn.f32 	%f1331, %f1361, %f1372, %f1331;
	fma.rn.f32 	%f1330, %f1362, %f1372, %f1330;
	fma.rn.f32 	%f1329, %f1363, %f1372, %f1329;
	fma.rn.f32 	%f1328, %f1364, %f1372, %f1328;
	fma.rn.f32 	%f1327, %f1365, %f1372, %f1327;
	fma.rn.f32 	%f1326, %f1366, %f1372, %f1326;
	setp.eq.s16	%p55, %rs161, 0;
	@%p55 bra 	BB0_45;

	div.rn.f32 	%f888, %f132, %f356;
	div.rn.f32 	%f889, %f888, %f96;
	cvt.sat.f32.f32	%f890, %f889;
	mul.f32 	%f1372, %f1372, %f890;

BB0_45:
	add.f32 	%f1325, %f1325, %f1372;

BB0_46:
	add.s32 	%r241, %r241, 1;
	setp.lt.u32	%p56, %r241, %r4;
	@%p56 bra 	BB0_5;

BB0_47:
	ld.global.u32 	%r247, [imageEnabled];
	and.b32  	%r116, %r247, 8;
	setp.eq.s32	%p57, %r116, 0;
	@%p57 bra 	BB0_60;

	cvt.sat.f32.f32	%f210, %f1325;
	cvt.u64.u32	%rd46, %r3;
	cvt.u64.u32	%rd45, %r2;
	mov.u64 	%rd49, image_Mask;
	cvta.global.u64 	%rd44, %rd49;
	// inline asm
	call (%rd43), _rt_buffer_get_64, (%rd44, %r28, %r28, %rd45, %rd46, %rd13, %rd13);
	// inline asm
	mov.f32 	%f893, 0f3E68BA2E;
	cvt.rzi.f32.f32	%f894, %f893;
	fma.rn.f32 	%f895, %f894, 0fC0000000, 0f3EE8BA2E;
	abs.f32 	%f211, %f895;
	abs.f32 	%f212, %f210;
	setp.lt.f32	%p58, %f212, 0f00800000;
	mul.f32 	%f896, %f212, 0f4B800000;
	selp.f32	%f897, 0fC3170000, 0fC2FE0000, %p58;
	selp.f32	%f898, %f896, %f212, %p58;
	mov.b32 	 %r119, %f898;
	and.b32  	%r120, %r119, 8388607;
	or.b32  	%r121, %r120, 1065353216;
	mov.b32 	 %f899, %r121;
	shr.u32 	%r122, %r119, 23;
	cvt.rn.f32.u32	%f900, %r122;
	add.f32 	%f901, %f897, %f900;
	setp.gt.f32	%p59, %f899, 0f3FB504F3;
	mul.f32 	%f902, %f899, 0f3F000000;
	add.f32 	%f903, %f901, 0f3F800000;
	selp.f32	%f904, %f902, %f899, %p59;
	selp.f32	%f905, %f903, %f901, %p59;
	add.f32 	%f906, %f904, 0fBF800000;
	add.f32 	%f892, %f904, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f891,%f892;
	// inline asm
	add.f32 	%f907, %f906, %f906;
	mul.f32 	%f908, %f891, %f907;
	mul.f32 	%f909, %f908, %f908;
	mov.f32 	%f910, 0f3C4CAF63;
	mov.f32 	%f911, 0f3B18F0FE;
	fma.rn.f32 	%f912, %f911, %f909, %f910;
	mov.f32 	%f913, 0f3DAAAABD;
	fma.rn.f32 	%f914, %f912, %f909, %f913;
	mul.rn.f32 	%f915, %f914, %f909;
	mul.rn.f32 	%f916, %f915, %f908;
	sub.f32 	%f917, %f906, %f908;
	neg.f32 	%f918, %f908;
	add.f32 	%f919, %f917, %f917;
	fma.rn.f32 	%f920, %f918, %f906, %f919;
	mul.rn.f32 	%f921, %f891, %f920;
	add.f32 	%f922, %f916, %f908;
	sub.f32 	%f923, %f908, %f922;
	add.f32 	%f924, %f916, %f923;
	add.f32 	%f925, %f921, %f924;
	add.f32 	%f926, %f922, %f925;
	sub.f32 	%f927, %f922, %f926;
	add.f32 	%f928, %f925, %f927;
	mov.f32 	%f929, 0f3F317200;
	mul.rn.f32 	%f930, %f905, %f929;
	mov.f32 	%f931, 0f35BFBE8E;
	mul.rn.f32 	%f932, %f905, %f931;
	add.f32 	%f933, %f930, %f926;
	sub.f32 	%f934, %f930, %f933;
	add.f32 	%f935, %f926, %f934;
	add.f32 	%f936, %f928, %f935;
	add.f32 	%f937, %f932, %f936;
	add.f32 	%f938, %f933, %f937;
	sub.f32 	%f939, %f933, %f938;
	add.f32 	%f940, %f937, %f939;
	mov.f32 	%f941, 0f3EE8BA2E;
	mul.rn.f32 	%f942, %f941, %f938;
	neg.f32 	%f943, %f942;
	fma.rn.f32 	%f944, %f941, %f938, %f943;
	fma.rn.f32 	%f945, %f941, %f940, %f944;
	mov.f32 	%f946, 0f00000000;
	fma.rn.f32 	%f947, %f946, %f938, %f945;
	add.rn.f32 	%f948, %f942, %f947;
	neg.f32 	%f949, %f948;
	add.rn.f32 	%f950, %f942, %f949;
	add.rn.f32 	%f951, %f950, %f947;
	mov.b32 	 %r123, %f948;
	setp.eq.s32	%p60, %r123, 1118925336;
	add.s32 	%r124, %r123, -1;
	mov.b32 	 %f952, %r124;
	add.f32 	%f953, %f951, 0f37000000;
	selp.f32	%f954, %f952, %f948, %p60;
	selp.f32	%f213, %f953, %f951, %p60;
	mul.f32 	%f955, %f954, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f956, %f955;
	mov.f32 	%f957, 0fBF317200;
	fma.rn.f32 	%f958, %f956, %f957, %f954;
	mov.f32 	%f959, 0fB5BFBE8E;
	fma.rn.f32 	%f960, %f956, %f959, %f958;
	mul.f32 	%f961, %f960, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f962, %f961;
	add.f32 	%f963, %f956, 0f00000000;
	ex2.approx.f32 	%f964, %f963;
	mul.f32 	%f965, %f962, %f964;
	setp.lt.f32	%p61, %f954, 0fC2D20000;
	selp.f32	%f966, 0f00000000, %f965, %p61;
	setp.gt.f32	%p62, %f954, 0f42D20000;
	selp.f32	%f1406, 0f7F800000, %f966, %p62;
	setp.eq.f32	%p63, %f1406, 0f7F800000;
	@%p63 bra 	BB0_50;

	fma.rn.f32 	%f1406, %f1406, %f213, %f1406;

BB0_50:
	setp.lt.f32	%p64, %f210, 0f00000000;
	setp.eq.f32	%p65, %f211, 0f3F800000;
	and.pred  	%p4, %p64, %p65;
	mov.b32 	 %r125, %f1406;
	xor.b32  	%r126, %r125, -2147483648;
	mov.b32 	 %f967, %r126;
	selp.f32	%f1408, %f967, %f1406, %p4;
	setp.eq.f32	%p66, %f210, 0f00000000;
	@%p66 bra 	BB0_53;
	bra.uni 	BB0_51;

BB0_53:
	add.f32 	%f970, %f210, %f210;
	selp.f32	%f1408, %f970, 0f00000000, %p65;
	bra.uni 	BB0_54;

BB0_118:
	mov.u64 	%rd190, image_HDR;
	cvta.global.u64 	%rd185, %rd190;
	mov.u32 	%r212, 8;
	// inline asm
	call (%rd184), _rt_buffer_get_64, (%rd185, %r28, %r212, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1290, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs107, %f1290;}

	// inline asm
	mov.u16 	%rs108, 0;
	st.v4.u16 	[%rd184], {%rs107, %rs107, %rs107, %rs108};

BB0_119:
	ld.global.u32 	%r213, [additive];
	setp.eq.s32	%p133, %r213, 0;
	@%p133 bra 	BB0_121;

	mov.u64 	%rd203, image_RNM0;
	cvta.global.u64 	%rd192, %rd203;
	mov.u32 	%r217, 8;
	// inline asm
	call (%rd191), _rt_buffer_get_64, (%rd192, %r28, %r217, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs115, %rs116, %rs117, %rs118}, [%rd191];
	// inline asm
	{  cvt.f32.f16 %f1291, %rs115;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1292, %rs116;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1293, %rs117;}

	// inline asm
	// inline asm
	call (%rd197), _rt_buffer_get_64, (%rd192, %r28, %r217, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1294, %f1291, 0f00000000;
	add.f32 	%f1295, %f1292, 0f00000000;
	add.f32 	%f1296, %f1293, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs114, %f1296;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs113, %f1295;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs112, %f1294;}

	// inline asm
	mov.u16 	%rs119, 0;
	st.v4.u16 	[%rd197], {%rs112, %rs113, %rs114, %rs119};
	bra.uni 	BB0_122;

BB0_121:
	mov.u64 	%rd210, image_RNM0;
	cvta.global.u64 	%rd205, %rd210;
	mov.u32 	%r219, 8;
	// inline asm
	call (%rd204), _rt_buffer_get_64, (%rd205, %r28, %r219, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1297, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs120, %f1297;}

	// inline asm
	mov.u16 	%rs121, 0;
	st.v4.u16 	[%rd204], {%rs120, %rs120, %rs120, %rs121};

BB0_122:
	ld.global.u32 	%r220, [additive];
	setp.eq.s32	%p134, %r220, 0;
	@%p134 bra 	BB0_124;

	mov.u64 	%rd223, image_RNM1;
	cvta.global.u64 	%rd212, %rd223;
	mov.u32 	%r224, 8;
	// inline asm
	call (%rd211), _rt_buffer_get_64, (%rd212, %r28, %r224, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs128, %rs129, %rs130, %rs131}, [%rd211];
	// inline asm
	{  cvt.f32.f16 %f1298, %rs128;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1299, %rs129;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1300, %rs130;}

	// inline asm
	// inline asm
	call (%rd217), _rt_buffer_get_64, (%rd212, %r28, %r224, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1301, %f1298, 0f00000000;
	add.f32 	%f1302, %f1299, 0f00000000;
	add.f32 	%f1303, %f1300, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs127, %f1303;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs126, %f1302;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs125, %f1301;}

	// inline asm
	mov.u16 	%rs132, 0;
	st.v4.u16 	[%rd217], {%rs125, %rs126, %rs127, %rs132};
	bra.uni 	BB0_125;

BB0_124:
	mov.u64 	%rd230, image_RNM1;
	cvta.global.u64 	%rd225, %rd230;
	mov.u32 	%r226, 8;
	// inline asm
	call (%rd224), _rt_buffer_get_64, (%rd225, %r28, %r226, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1304, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs133, %f1304;}

	// inline asm
	mov.u16 	%rs134, 0;
	st.v4.u16 	[%rd224], {%rs133, %rs133, %rs133, %rs134};

BB0_125:
	ld.global.u32 	%r227, [additive];
	setp.eq.s32	%p135, %r227, 0;
	@%p135 bra 	BB0_127;

	mov.u64 	%rd243, image_RNM2;
	cvta.global.u64 	%rd232, %rd243;
	mov.u32 	%r231, 8;
	// inline asm
	call (%rd231), _rt_buffer_get_64, (%rd232, %r28, %r231, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs141, %rs142, %rs143, %rs144}, [%rd231];
	// inline asm
	{  cvt.f32.f16 %f1305, %rs141;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1306, %rs142;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1307, %rs143;}

	// inline asm
	// inline asm
	call (%rd237), _rt_buffer_get_64, (%rd232, %r28, %r231, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1308, %f1305, 0f00000000;
	add.f32 	%f1309, %f1306, 0f00000000;
	add.f32 	%f1310, %f1307, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs140, %f1310;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs139, %f1309;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs138, %f1308;}

	// inline asm
	mov.u16 	%rs145, 0;
	st.v4.u16 	[%rd237], {%rs138, %rs139, %rs140, %rs145};
	bra.uni 	BB0_128;

BB0_127:
	mov.u64 	%rd250, image_RNM2;
	cvta.global.u64 	%rd245, %rd250;
	mov.u32 	%r233, 8;
	// inline asm
	call (%rd244), _rt_buffer_get_64, (%rd245, %r28, %r233, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1311, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs146, %f1311;}

	// inline asm
	mov.u16 	%rs147, 0;
	st.v4.u16 	[%rd244], {%rs146, %rs146, %rs146, %rs147};

BB0_128:
	ld.global.u32 	%r234, [additive];
	setp.eq.s32	%p136, %r234, 0;
	@%p136 bra 	BB0_130;

	mov.u64 	%rd263, image_RNM3;
	cvta.global.u64 	%rd252, %rd263;
	mov.u32 	%r238, 8;
	// inline asm
	call (%rd251), _rt_buffer_get_64, (%rd252, %r28, %r238, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs154, %rs155, %rs156, %rs157}, [%rd251];
	// inline asm
	{  cvt.f32.f16 %f1312, %rs154;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1313, %rs155;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1314, %rs156;}

	// inline asm
	// inline asm
	call (%rd257), _rt_buffer_get_64, (%rd252, %r28, %r238, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1315, %f1312, 0f00000000;
	add.f32 	%f1316, %f1313, 0f00000000;
	add.f32 	%f1317, %f1314, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs153, %f1317;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs152, %f1316;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs151, %f1315;}

	// inline asm
	mov.u16 	%rs158, 0;
	st.v4.u16 	[%rd257], {%rs151, %rs152, %rs153, %rs158};
	bra.uni 	BB0_131;

BB0_130:
	mov.u64 	%rd270, image_RNM3;
	cvta.global.u64 	%rd265, %rd270;
	mov.u32 	%r240, 8;
	// inline asm
	call (%rd264), _rt_buffer_get_64, (%rd265, %r28, %r240, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1318, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs159, %f1318;}

	// inline asm
	mov.u16 	%rs160, 0;
	st.v4.u16 	[%rd264], {%rs159, %rs159, %rs159, %rs160};
	bra.uni 	BB0_131;

BB0_51:
	setp.geu.f32	%p67, %f210, 0f00000000;
	@%p67 bra 	BB0_54;

	cvt.rzi.f32.f32	%f969, %f941;
	setp.neu.f32	%p68, %f969, 0f3EE8BA2E;
	selp.f32	%f1408, 0f7FFFFFFF, %f1408, %p68;

BB0_54:
	add.f32 	%f971, %f212, 0f3EE8BA2E;
	mov.b32 	 %r127, %f971;
	setp.lt.s32	%p70, %r127, 2139095040;
	@%p70 bra 	BB0_59;

	setp.gtu.f32	%p71, %f212, 0f7F800000;
	@%p71 bra 	BB0_58;
	bra.uni 	BB0_56;

BB0_58:
	add.f32 	%f1408, %f210, 0f3EE8BA2E;
	bra.uni 	BB0_59;

BB0_56:
	setp.neu.f32	%p72, %f212, 0f7F800000;
	@%p72 bra 	BB0_59;

	selp.f32	%f1408, 0fFF800000, 0f7F800000, %p4;

BB0_59:
	mul.f32 	%f972, %f1408, 0f437F0000;
	setp.eq.f32	%p73, %f210, 0f3F800000;
	selp.f32	%f973, 0f437F0000, %f972, %p73;
	cvt.rzi.u32.f32	%r128, %f973;
	cvt.u16.u32	%rs17, %r128;
	mov.u16 	%rs18, 255;
	st.v2.u8 	[%rd43], {%rs17, %rs18};
	ld.global.u32 	%r247, [imageEnabled];

BB0_60:
	and.b32  	%r129, %r247, 1;
	setp.eq.b32	%p74, %r129, 1;
	@!%p74 bra 	BB0_95;
	bra.uni 	BB0_61;

BB0_61:
	mov.f32 	%f976, 0f3E666666;
	cvt.rzi.f32.f32	%f977, %f976;
	fma.rn.f32 	%f978, %f977, 0fC0000000, 0f3EE66666;
	abs.f32 	%f224, %f978;
	abs.f32 	%f225, %f1340;
	setp.lt.f32	%p75, %f225, 0f00800000;
	mul.f32 	%f979, %f225, 0f4B800000;
	selp.f32	%f980, 0fC3170000, 0fC2FE0000, %p75;
	selp.f32	%f981, %f979, %f225, %p75;
	mov.b32 	 %r130, %f981;
	and.b32  	%r131, %r130, 8388607;
	or.b32  	%r132, %r131, 1065353216;
	mov.b32 	 %f982, %r132;
	shr.u32 	%r133, %r130, 23;
	cvt.rn.f32.u32	%f983, %r133;
	add.f32 	%f984, %f980, %f983;
	setp.gt.f32	%p76, %f982, 0f3FB504F3;
	mul.f32 	%f985, %f982, 0f3F000000;
	add.f32 	%f986, %f984, 0f3F800000;
	selp.f32	%f987, %f985, %f982, %p76;
	selp.f32	%f988, %f986, %f984, %p76;
	add.f32 	%f989, %f987, 0fBF800000;
	add.f32 	%f975, %f987, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f974,%f975;
	// inline asm
	add.f32 	%f990, %f989, %f989;
	mul.f32 	%f991, %f974, %f990;
	mul.f32 	%f992, %f991, %f991;
	mov.f32 	%f993, 0f3C4CAF63;
	mov.f32 	%f994, 0f3B18F0FE;
	fma.rn.f32 	%f995, %f994, %f992, %f993;
	mov.f32 	%f996, 0f3DAAAABD;
	fma.rn.f32 	%f997, %f995, %f992, %f996;
	mul.rn.f32 	%f998, %f997, %f992;
	mul.rn.f32 	%f999, %f998, %f991;
	sub.f32 	%f1000, %f989, %f991;
	neg.f32 	%f1001, %f991;
	add.f32 	%f1002, %f1000, %f1000;
	fma.rn.f32 	%f1003, %f1001, %f989, %f1002;
	mul.rn.f32 	%f1004, %f974, %f1003;
	add.f32 	%f1005, %f999, %f991;
	sub.f32 	%f1006, %f991, %f1005;
	add.f32 	%f1007, %f999, %f1006;
	add.f32 	%f1008, %f1004, %f1007;
	add.f32 	%f1009, %f1005, %f1008;
	sub.f32 	%f1010, %f1005, %f1009;
	add.f32 	%f1011, %f1008, %f1010;
	mov.f32 	%f1012, 0f3F317200;
	mul.rn.f32 	%f1013, %f988, %f1012;
	mov.f32 	%f1014, 0f35BFBE8E;
	mul.rn.f32 	%f1015, %f988, %f1014;
	add.f32 	%f1016, %f1013, %f1009;
	sub.f32 	%f1017, %f1013, %f1016;
	add.f32 	%f1018, %f1009, %f1017;
	add.f32 	%f1019, %f1011, %f1018;
	add.f32 	%f1020, %f1015, %f1019;
	add.f32 	%f1021, %f1016, %f1020;
	sub.f32 	%f1022, %f1016, %f1021;
	add.f32 	%f1023, %f1020, %f1022;
	mov.f32 	%f1024, 0f3EE66666;
	mul.rn.f32 	%f1025, %f1024, %f1021;
	neg.f32 	%f1026, %f1025;
	fma.rn.f32 	%f1027, %f1024, %f1021, %f1026;
	fma.rn.f32 	%f1028, %f1024, %f1023, %f1027;
	mov.f32 	%f1029, 0f00000000;
	fma.rn.f32 	%f1030, %f1029, %f1021, %f1028;
	add.rn.f32 	%f1031, %f1025, %f1030;
	neg.f32 	%f1032, %f1031;
	add.rn.f32 	%f1033, %f1025, %f1032;
	add.rn.f32 	%f1034, %f1033, %f1030;
	mov.b32 	 %r134, %f1031;
	setp.eq.s32	%p77, %r134, 1118925336;
	add.s32 	%r135, %r134, -1;
	mov.b32 	 %f1035, %r135;
	add.f32 	%f1036, %f1034, 0f37000000;
	selp.f32	%f1037, %f1035, %f1031, %p77;
	selp.f32	%f226, %f1036, %f1034, %p77;
	mul.f32 	%f1038, %f1037, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1039, %f1038;
	mov.f32 	%f1040, 0fBF317200;
	fma.rn.f32 	%f1041, %f1039, %f1040, %f1037;
	mov.f32 	%f1042, 0fB5BFBE8E;
	fma.rn.f32 	%f1043, %f1039, %f1042, %f1041;
	mul.f32 	%f1044, %f1043, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1045, %f1044;
	add.f32 	%f1046, %f1039, 0f00000000;
	ex2.approx.f32 	%f1047, %f1046;
	mul.f32 	%f1048, %f1045, %f1047;
	setp.lt.f32	%p78, %f1037, 0fC2D20000;
	selp.f32	%f1049, 0f00000000, %f1048, %p78;
	setp.gt.f32	%p79, %f1037, 0f42D20000;
	selp.f32	%f1409, 0f7F800000, %f1049, %p79;
	setp.eq.f32	%p80, %f1409, 0f7F800000;
	@%p80 bra 	BB0_63;

	fma.rn.f32 	%f1409, %f1409, %f226, %f1409;

BB0_63:
	setp.lt.f32	%p81, %f1340, 0f00000000;
	setp.eq.f32	%p82, %f224, 0f3F800000;
	and.pred  	%p5, %p81, %p82;
	mov.b32 	 %r136, %f1409;
	xor.b32  	%r137, %r136, -2147483648;
	mov.b32 	 %f1050, %r137;
	selp.f32	%f1411, %f1050, %f1409, %p5;
	setp.eq.f32	%p83, %f1340, 0f00000000;
	@%p83 bra 	BB0_66;
	bra.uni 	BB0_64;

BB0_66:
	add.f32 	%f1053, %f1340, %f1340;
	selp.f32	%f1411, %f1053, 0f00000000, %p82;
	bra.uni 	BB0_67;

BB0_64:
	setp.geu.f32	%p84, %f1340, 0f00000000;
	@%p84 bra 	BB0_67;

	cvt.rzi.f32.f32	%f1052, %f1024;
	setp.neu.f32	%p85, %f1052, 0f3EE66666;
	selp.f32	%f1411, 0f7FFFFFFF, %f1411, %p85;

BB0_67:
	add.f32 	%f1054, %f225, 0f3EE66666;
	mov.b32 	 %r138, %f1054;
	setp.lt.s32	%p87, %r138, 2139095040;
	@%p87 bra 	BB0_72;

	setp.gtu.f32	%p88, %f225, 0f7F800000;
	@%p88 bra 	BB0_71;
	bra.uni 	BB0_69;

BB0_71:
	add.f32 	%f1411, %f1340, 0f3EE66666;
	bra.uni 	BB0_72;

BB0_69:
	setp.neu.f32	%p89, %f225, 0f7F800000;
	@%p89 bra 	BB0_72;

	selp.f32	%f1411, 0fFF800000, 0f7F800000, %p5;

BB0_72:
	setp.eq.f32	%p90, %f1340, 0f3F800000;
	selp.f32	%f237, 0f3F800000, %f1411, %p90;
	abs.f32 	%f238, %f1339;
	setp.lt.f32	%p91, %f238, 0f00800000;
	mul.f32 	%f1057, %f238, 0f4B800000;
	selp.f32	%f1058, 0fC3170000, 0fC2FE0000, %p91;
	selp.f32	%f1059, %f1057, %f238, %p91;
	mov.b32 	 %r139, %f1059;
	and.b32  	%r140, %r139, 8388607;
	or.b32  	%r141, %r140, 1065353216;
	mov.b32 	 %f1060, %r141;
	shr.u32 	%r142, %r139, 23;
	cvt.rn.f32.u32	%f1061, %r142;
	add.f32 	%f1062, %f1058, %f1061;
	setp.gt.f32	%p92, %f1060, 0f3FB504F3;
	mul.f32 	%f1063, %f1060, 0f3F000000;
	add.f32 	%f1064, %f1062, 0f3F800000;
	selp.f32	%f1065, %f1063, %f1060, %p92;
	selp.f32	%f1066, %f1064, %f1062, %p92;
	add.f32 	%f1067, %f1065, 0fBF800000;
	add.f32 	%f1056, %f1065, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1055,%f1056;
	// inline asm
	add.f32 	%f1068, %f1067, %f1067;
	mul.f32 	%f1069, %f1055, %f1068;
	mul.f32 	%f1070, %f1069, %f1069;
	fma.rn.f32 	%f1073, %f994, %f1070, %f993;
	fma.rn.f32 	%f1075, %f1073, %f1070, %f996;
	mul.rn.f32 	%f1076, %f1075, %f1070;
	mul.rn.f32 	%f1077, %f1076, %f1069;
	sub.f32 	%f1078, %f1067, %f1069;
	neg.f32 	%f1079, %f1069;
	add.f32 	%f1080, %f1078, %f1078;
	fma.rn.f32 	%f1081, %f1079, %f1067, %f1080;
	mul.rn.f32 	%f1082, %f1055, %f1081;
	add.f32 	%f1083, %f1077, %f1069;
	sub.f32 	%f1084, %f1069, %f1083;
	add.f32 	%f1085, %f1077, %f1084;
	add.f32 	%f1086, %f1082, %f1085;
	add.f32 	%f1087, %f1083, %f1086;
	sub.f32 	%f1088, %f1083, %f1087;
	add.f32 	%f1089, %f1086, %f1088;
	mul.rn.f32 	%f1091, %f1066, %f1012;
	mul.rn.f32 	%f1093, %f1066, %f1014;
	add.f32 	%f1094, %f1091, %f1087;
	sub.f32 	%f1095, %f1091, %f1094;
	add.f32 	%f1096, %f1087, %f1095;
	add.f32 	%f1097, %f1089, %f1096;
	add.f32 	%f1098, %f1093, %f1097;
	add.f32 	%f1099, %f1094, %f1098;
	sub.f32 	%f1100, %f1094, %f1099;
	add.f32 	%f1101, %f1098, %f1100;
	mul.rn.f32 	%f1103, %f1024, %f1099;
	neg.f32 	%f1104, %f1103;
	fma.rn.f32 	%f1105, %f1024, %f1099, %f1104;
	fma.rn.f32 	%f1106, %f1024, %f1101, %f1105;
	fma.rn.f32 	%f1108, %f1029, %f1099, %f1106;
	add.rn.f32 	%f1109, %f1103, %f1108;
	neg.f32 	%f1110, %f1109;
	add.rn.f32 	%f1111, %f1103, %f1110;
	add.rn.f32 	%f1112, %f1111, %f1108;
	mov.b32 	 %r143, %f1109;
	setp.eq.s32	%p93, %r143, 1118925336;
	add.s32 	%r144, %r143, -1;
	mov.b32 	 %f1113, %r144;
	add.f32 	%f1114, %f1112, 0f37000000;
	selp.f32	%f1115, %f1113, %f1109, %p93;
	selp.f32	%f239, %f1114, %f1112, %p93;
	mul.f32 	%f1116, %f1115, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1117, %f1116;
	fma.rn.f32 	%f1119, %f1117, %f1040, %f1115;
	fma.rn.f32 	%f1121, %f1117, %f1042, %f1119;
	mul.f32 	%f1122, %f1121, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1123, %f1122;
	add.f32 	%f1124, %f1117, 0f00000000;
	ex2.approx.f32 	%f1125, %f1124;
	mul.f32 	%f1126, %f1123, %f1125;
	setp.lt.f32	%p94, %f1115, 0fC2D20000;
	selp.f32	%f1127, 0f00000000, %f1126, %p94;
	setp.gt.f32	%p95, %f1115, 0f42D20000;
	selp.f32	%f1412, 0f7F800000, %f1127, %p95;
	setp.eq.f32	%p96, %f1412, 0f7F800000;
	@%p96 bra 	BB0_74;

	fma.rn.f32 	%f1412, %f1412, %f239, %f1412;

BB0_74:
	setp.lt.f32	%p97, %f1339, 0f00000000;
	and.pred  	%p6, %p97, %p82;
	mov.b32 	 %r145, %f1412;
	xor.b32  	%r146, %r145, -2147483648;
	mov.b32 	 %f1128, %r146;
	selp.f32	%f1414, %f1128, %f1412, %p6;
	setp.eq.f32	%p99, %f1339, 0f00000000;
	@%p99 bra 	BB0_77;
	bra.uni 	BB0_75;

BB0_77:
	add.f32 	%f1131, %f1339, %f1339;
	selp.f32	%f1414, %f1131, 0f00000000, %p82;
	bra.uni 	BB0_78;

BB0_75:
	setp.geu.f32	%p100, %f1339, 0f00000000;
	@%p100 bra 	BB0_78;

	cvt.rzi.f32.f32	%f1130, %f1024;
	setp.neu.f32	%p101, %f1130, 0f3EE66666;
	selp.f32	%f1414, 0f7FFFFFFF, %f1414, %p101;

BB0_78:
	add.f32 	%f1132, %f238, 0f3EE66666;
	mov.b32 	 %r147, %f1132;
	setp.lt.s32	%p103, %r147, 2139095040;
	@%p103 bra 	BB0_83;

	setp.gtu.f32	%p104, %f238, 0f7F800000;
	@%p104 bra 	BB0_82;
	bra.uni 	BB0_80;

BB0_82:
	add.f32 	%f1414, %f1339, 0f3EE66666;
	bra.uni 	BB0_83;

BB0_80:
	setp.neu.f32	%p105, %f238, 0f7F800000;
	@%p105 bra 	BB0_83;

	selp.f32	%f1414, 0fFF800000, 0f7F800000, %p6;

BB0_83:
	setp.eq.f32	%p106, %f1339, 0f3F800000;
	selp.f32	%f250, 0f3F800000, %f1414, %p106;
	abs.f32 	%f251, %f1338;
	setp.lt.f32	%p107, %f251, 0f00800000;
	mul.f32 	%f1135, %f251, 0f4B800000;
	selp.f32	%f1136, 0fC3170000, 0fC2FE0000, %p107;
	selp.f32	%f1137, %f1135, %f251, %p107;
	mov.b32 	 %r148, %f1137;
	and.b32  	%r149, %r148, 8388607;
	or.b32  	%r150, %r149, 1065353216;
	mov.b32 	 %f1138, %r150;
	shr.u32 	%r151, %r148, 23;
	cvt.rn.f32.u32	%f1139, %r151;
	add.f32 	%f1140, %f1136, %f1139;
	setp.gt.f32	%p108, %f1138, 0f3FB504F3;
	mul.f32 	%f1141, %f1138, 0f3F000000;
	add.f32 	%f1142, %f1140, 0f3F800000;
	selp.f32	%f1143, %f1141, %f1138, %p108;
	selp.f32	%f1144, %f1142, %f1140, %p108;
	add.f32 	%f1145, %f1143, 0fBF800000;
	add.f32 	%f1134, %f1143, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1133,%f1134;
	// inline asm
	add.f32 	%f1146, %f1145, %f1145;
	mul.f32 	%f1147, %f1133, %f1146;
	mul.f32 	%f1148, %f1147, %f1147;
	fma.rn.f32 	%f1151, %f994, %f1148, %f993;
	fma.rn.f32 	%f1153, %f1151, %f1148, %f996;
	mul.rn.f32 	%f1154, %f1153, %f1148;
	mul.rn.f32 	%f1155, %f1154, %f1147;
	sub.f32 	%f1156, %f1145, %f1147;
	neg.f32 	%f1157, %f1147;
	add.f32 	%f1158, %f1156, %f1156;
	fma.rn.f32 	%f1159, %f1157, %f1145, %f1158;
	mul.rn.f32 	%f1160, %f1133, %f1159;
	add.f32 	%f1161, %f1155, %f1147;
	sub.f32 	%f1162, %f1147, %f1161;
	add.f32 	%f1163, %f1155, %f1162;
	add.f32 	%f1164, %f1160, %f1163;
	add.f32 	%f1165, %f1161, %f1164;
	sub.f32 	%f1166, %f1161, %f1165;
	add.f32 	%f1167, %f1164, %f1166;
	mul.rn.f32 	%f1169, %f1144, %f1012;
	mul.rn.f32 	%f1171, %f1144, %f1014;
	add.f32 	%f1172, %f1169, %f1165;
	sub.f32 	%f1173, %f1169, %f1172;
	add.f32 	%f1174, %f1165, %f1173;
	add.f32 	%f1175, %f1167, %f1174;
	add.f32 	%f1176, %f1171, %f1175;
	add.f32 	%f1177, %f1172, %f1176;
	sub.f32 	%f1178, %f1172, %f1177;
	add.f32 	%f1179, %f1176, %f1178;
	mul.rn.f32 	%f1181, %f1024, %f1177;
	neg.f32 	%f1182, %f1181;
	fma.rn.f32 	%f1183, %f1024, %f1177, %f1182;
	fma.rn.f32 	%f1184, %f1024, %f1179, %f1183;
	fma.rn.f32 	%f1186, %f1029, %f1177, %f1184;
	add.rn.f32 	%f1187, %f1181, %f1186;
	neg.f32 	%f1188, %f1187;
	add.rn.f32 	%f1189, %f1181, %f1188;
	add.rn.f32 	%f1190, %f1189, %f1186;
	mov.b32 	 %r152, %f1187;
	setp.eq.s32	%p109, %r152, 1118925336;
	add.s32 	%r153, %r152, -1;
	mov.b32 	 %f1191, %r153;
	add.f32 	%f1192, %f1190, 0f37000000;
	selp.f32	%f1193, %f1191, %f1187, %p109;
	selp.f32	%f252, %f1192, %f1190, %p109;
	mul.f32 	%f1194, %f1193, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1195, %f1194;
	fma.rn.f32 	%f1197, %f1195, %f1040, %f1193;
	fma.rn.f32 	%f1199, %f1195, %f1042, %f1197;
	mul.f32 	%f1200, %f1199, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1201, %f1200;
	add.f32 	%f1202, %f1195, 0f00000000;
	ex2.approx.f32 	%f1203, %f1202;
	mul.f32 	%f1204, %f1201, %f1203;
	setp.lt.f32	%p110, %f1193, 0fC2D20000;
	selp.f32	%f1205, 0f00000000, %f1204, %p110;
	setp.gt.f32	%p111, %f1193, 0f42D20000;
	selp.f32	%f1415, 0f7F800000, %f1205, %p111;
	setp.eq.f32	%p112, %f1415, 0f7F800000;
	@%p112 bra 	BB0_85;

	fma.rn.f32 	%f1415, %f1415, %f252, %f1415;

BB0_85:
	setp.lt.f32	%p113, %f1338, 0f00000000;
	and.pred  	%p7, %p113, %p82;
	mov.b32 	 %r154, %f1415;
	xor.b32  	%r155, %r154, -2147483648;
	mov.b32 	 %f1206, %r155;
	selp.f32	%f1417, %f1206, %f1415, %p7;
	setp.eq.f32	%p115, %f1338, 0f00000000;
	@%p115 bra 	BB0_88;
	bra.uni 	BB0_86;

BB0_88:
	add.f32 	%f1209, %f1338, %f1338;
	selp.f32	%f1417, %f1209, 0f00000000, %p82;
	bra.uni 	BB0_89;

BB0_86:
	setp.geu.f32	%p116, %f1338, 0f00000000;
	@%p116 bra 	BB0_89;

	cvt.rzi.f32.f32	%f1208, %f1024;
	setp.neu.f32	%p117, %f1208, 0f3EE66666;
	selp.f32	%f1417, 0f7FFFFFFF, %f1417, %p117;

BB0_89:
	add.f32 	%f1210, %f251, 0f3EE66666;
	mov.b32 	 %r156, %f1210;
	setp.lt.s32	%p119, %r156, 2139095040;
	@%p119 bra 	BB0_94;

	setp.gtu.f32	%p120, %f251, 0f7F800000;
	@%p120 bra 	BB0_93;
	bra.uni 	BB0_91;

BB0_93:
	add.f32 	%f1417, %f1338, 0f3EE66666;
	bra.uni 	BB0_94;

BB0_91:
	setp.neu.f32	%p121, %f251, 0f7F800000;
	@%p121 bra 	BB0_94;

	selp.f32	%f1417, 0fFF800000, 0f7F800000, %p7;

BB0_94:
	setp.eq.f32	%p122, %f1338, 0f3F800000;
	selp.f32	%f1211, 0f3F800000, %f1417, %p122;
	cvt.u64.u32	%rd53, %r3;
	cvt.u64.u32	%rd52, %r2;
	mov.u64 	%rd56, image;
	cvta.global.u64 	%rd51, %rd56;
	// inline asm
	call (%rd50), _rt_buffer_get_64, (%rd51, %r28, %r29, %rd52, %rd53, %rd13, %rd13);
	// inline asm
	cvt.sat.f32.f32	%f1212, %f1211;
	mul.f32 	%f1213, %f1212, 0f437FFD71;
	cvt.rzi.u32.f32	%r159, %f1213;
	cvt.sat.f32.f32	%f1214, %f250;
	mul.f32 	%f1215, %f1214, 0f437FFD71;
	cvt.rzi.u32.f32	%r160, %f1215;
	cvt.sat.f32.f32	%f1216, %f237;
	mul.f32 	%f1217, %f1216, 0f437FFD71;
	cvt.rzi.u32.f32	%r161, %f1217;
	cvt.u16.u32	%rs19, %r159;
	cvt.u16.u32	%rs20, %r161;
	cvt.u16.u32	%rs21, %r160;
	mov.u16 	%rs22, 255;
	st.v4.u8 	[%rd50], {%rs19, %rs21, %rs20, %rs22};
	ld.global.u32 	%r247, [imageEnabled];

BB0_95:
	cvt.u64.u32	%rd4, %r2;
	cvt.u64.u32	%rd5, %r3;
	and.b32  	%r162, %r247, 4;
	setp.eq.s32	%p123, %r162, 0;
	@%p123 bra 	BB0_99;

	ld.global.u32 	%r163, [additive];
	setp.eq.s32	%p124, %r163, 0;
	mov.f32 	%f1218, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs23, %f1218;}

	// inline asm
	@%p124 bra 	BB0_98;

	mov.u64 	%rd69, image_HDR;
	cvta.global.u64 	%rd58, %rd69;
	mov.u32 	%r167, 8;
	// inline asm
	call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r167, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs30, %rs31, %rs32, %rs33}, [%rd57];
	// inline asm
	{  cvt.f32.f16 %f1219, %rs30;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1220, %rs31;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1221, %rs32;}

	// inline asm
	// inline asm
	call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r167, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1222, %f1340, %f1219;
	add.f32 	%f1223, %f1339, %f1220;
	add.f32 	%f1224, %f1338, %f1221;
	// inline asm
	{  cvt.rn.f16.f32 %rs29, %f1224;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs28, %f1223;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs27, %f1222;}

	// inline asm
	st.v4.u16 	[%rd63], {%rs27, %rs28, %rs29, %rs23};
	bra.uni 	BB0_99;

BB0_98:
	mov.u64 	%rd76, image_HDR;
	cvta.global.u64 	%rd71, %rd76;
	mov.u32 	%r169, 8;
	// inline asm
	call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r169, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs36, %f1338;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs35, %f1339;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs34, %f1340;}

	// inline asm
	st.v4.u16 	[%rd70], {%rs34, %rs35, %rs36, %rs23};

BB0_99:
	mov.f32 	%f1229, 0f34000000;
	max.f32 	%f1230, %f1337, %f1229;
	div.rn.f32 	%f1231, %f1334, %f1230;
	max.f32 	%f1232, %f1336, %f1229;
	div.rn.f32 	%f1233, %f1333, %f1232;
	max.f32 	%f1234, %f1335, %f1229;
	div.rn.f32 	%f1235, %f1332, %f1234;
	fma.rn.f32 	%f263, %f1231, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f264, %f1233, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f265, %f1235, 0f3F000000, 0f3F000000;
	div.rn.f32 	%f1236, %f1331, %f1230;
	div.rn.f32 	%f1237, %f1330, %f1232;
	div.rn.f32 	%f1238, %f1329, %f1234;
	fma.rn.f32 	%f266, %f1236, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f267, %f1237, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f268, %f1238, 0f3F000000, 0f3F000000;
	div.rn.f32 	%f1239, %f1328, %f1230;
	div.rn.f32 	%f1240, %f1327, %f1232;
	div.rn.f32 	%f1241, %f1326, %f1234;
	fma.rn.f32 	%f269, %f1239, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f270, %f1240, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f271, %f1241, 0f3F000000, 0f3F000000;
	ld.global.u32 	%r170, [additive];
	setp.eq.s32	%p125, %r170, 0;
	mov.f32 	%f1228, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs37, %f1228;}

	// inline asm
	@%p125 bra 	BB0_101;

	mov.u64 	%rd89, image_RNM0;
	cvta.global.u64 	%rd78, %rd89;
	mov.u32 	%r174, 8;
	// inline asm
	call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r174, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs44, %rs45, %rs46, %rs47}, [%rd77];
	// inline asm
	{  cvt.f32.f16 %f1242, %rs44;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1243, %rs45;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1244, %rs46;}

	// inline asm
	// inline asm
	call (%rd83), _rt_buffer_get_64, (%rd78, %r28, %r174, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1245, %f1337, %f1242;
	add.f32 	%f1246, %f1336, %f1243;
	add.f32 	%f1247, %f1335, %f1244;
	// inline asm
	{  cvt.rn.f16.f32 %rs43, %f1247;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs42, %f1246;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs41, %f1245;}

	// inline asm
	st.v4.u16 	[%rd83], {%rs41, %rs42, %rs43, %rs37};
	bra.uni 	BB0_102;

BB0_101:
	mov.u64 	%rd96, image_RNM0;
	cvta.global.u64 	%rd91, %rd96;
	mov.u32 	%r176, 8;
	// inline asm
	call (%rd90), _rt_buffer_get_64, (%rd91, %r28, %r176, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs50, %f1335;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs49, %f1336;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs48, %f1337;}

	// inline asm
	st.v4.u16 	[%rd90], {%rs48, %rs49, %rs50, %rs37};

BB0_102:
	ld.global.u32 	%r177, [additive];
	setp.eq.s32	%p126, %r177, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs51, %f1228;}

	// inline asm
	@%p126 bra 	BB0_104;

	mov.u64 	%rd109, image_RNM1;
	cvta.global.u64 	%rd98, %rd109;
	mov.u32 	%r181, 8;
	// inline asm
	call (%rd97), _rt_buffer_get_64, (%rd98, %r28, %r181, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs58, %rs59, %rs60, %rs61}, [%rd97];
	// inline asm
	{  cvt.f32.f16 %f1252, %rs58;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1253, %rs59;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1254, %rs60;}

	// inline asm
	// inline asm
	call (%rd103), _rt_buffer_get_64, (%rd98, %r28, %r181, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1255, %f263, %f1252;
	add.f32 	%f1256, %f264, %f1253;
	add.f32 	%f1257, %f265, %f1254;
	// inline asm
	{  cvt.rn.f16.f32 %rs57, %f1257;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs56, %f1256;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs55, %f1255;}

	// inline asm
	st.v4.u16 	[%rd103], {%rs55, %rs56, %rs57, %rs51};
	bra.uni 	BB0_105;

BB0_104:
	mov.u64 	%rd116, image_RNM1;
	cvta.global.u64 	%rd111, %rd116;
	mov.u32 	%r183, 8;
	// inline asm
	call (%rd110), _rt_buffer_get_64, (%rd111, %r28, %r183, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs64, %f265;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs63, %f264;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs62, %f263;}

	// inline asm
	st.v4.u16 	[%rd110], {%rs62, %rs63, %rs64, %rs51};

BB0_105:
	ld.global.u32 	%r184, [additive];
	setp.eq.s32	%p127, %r184, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs65, %f1228;}

	// inline asm
	@%p127 bra 	BB0_107;

	mov.u64 	%rd129, image_RNM2;
	cvta.global.u64 	%rd118, %rd129;
	mov.u32 	%r188, 8;
	// inline asm
	call (%rd117), _rt_buffer_get_64, (%rd118, %r28, %r188, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs72, %rs73, %rs74, %rs75}, [%rd117];
	// inline asm
	{  cvt.f32.f16 %f1262, %rs72;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1263, %rs73;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1264, %rs74;}

	// inline asm
	// inline asm
	call (%rd123), _rt_buffer_get_64, (%rd118, %r28, %r188, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1265, %f266, %f1262;
	add.f32 	%f1266, %f267, %f1263;
	add.f32 	%f1267, %f268, %f1264;
	// inline asm
	{  cvt.rn.f16.f32 %rs71, %f1267;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs70, %f1266;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs69, %f1265;}

	// inline asm
	st.v4.u16 	[%rd123], {%rs69, %rs70, %rs71, %rs65};
	bra.uni 	BB0_108;

BB0_107:
	mov.u64 	%rd136, image_RNM2;
	cvta.global.u64 	%rd131, %rd136;
	mov.u32 	%r190, 8;
	// inline asm
	call (%rd130), _rt_buffer_get_64, (%rd131, %r28, %r190, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs78, %f268;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs77, %f267;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs76, %f266;}

	// inline asm
	st.v4.u16 	[%rd130], {%rs76, %rs77, %rs78, %rs65};

BB0_108:
	ld.global.u32 	%r191, [additive];
	setp.eq.s32	%p128, %r191, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs79, %f1228;}

	// inline asm
	@%p128 bra 	BB0_110;

	mov.u64 	%rd149, image_RNM3;
	cvta.global.u64 	%rd138, %rd149;
	mov.u32 	%r195, 8;
	// inline asm
	call (%rd137), _rt_buffer_get_64, (%rd138, %r28, %r195, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs86, %rs87, %rs88, %rs89}, [%rd137];
	// inline asm
	{  cvt.f32.f16 %f1272, %rs86;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1273, %rs87;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1274, %rs88;}

	// inline asm
	// inline asm
	call (%rd143), _rt_buffer_get_64, (%rd138, %r28, %r195, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1275, %f269, %f1272;
	add.f32 	%f1276, %f270, %f1273;
	add.f32 	%f1277, %f271, %f1274;
	// inline asm
	{  cvt.rn.f16.f32 %rs85, %f1277;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs84, %f1276;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs83, %f1275;}

	// inline asm
	st.v4.u16 	[%rd143], {%rs83, %rs84, %rs85, %rs79};
	bra.uni 	BB0_131;

BB0_110:
	mov.u64 	%rd156, image_RNM3;
	cvta.global.u64 	%rd151, %rd156;
	mov.u32 	%r197, 8;
	// inline asm
	call (%rd150), _rt_buffer_get_64, (%rd151, %r28, %r197, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs92, %f271;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs91, %f270;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs90, %f269;}

	// inline asm
	st.v4.u16 	[%rd150], {%rs90, %rs91, %rs92, %rs79};

BB0_131:
	ret;
}